Encode structured Fields without allocations, byte-equal to encoding/json

[eitamring]

Last code PR of the audit series (#47, #48). The structured output path (records carrying decoded Fields) went through encoding/json at ~1.6us and 10 allocs per record; the plain path was already zero-alloc.

• Hand-rolled fast path for the value shapes the pipeline produces (string, json.Number, bool, nil, nested maps and arrays), with sorted keys and escaping matching encoding/json exactly. Measured: 732-794ns, 0 B/op, 0 allocs (count=5); plain-record path unchanged.
• Anything the fast path cannot prove equivalent rewinds and falls back to encoding/json: unsupported types, non-finite floats, malformed json.Number, and nesting past a depth bound of 64. The bound counts arrays as well as objects, so cyclic maps and cyclic slices reach the fallback instead of overflowing the stack; deep acyclic values encode identically, cycles return the reference error.
• Key buffers release their string references on every exit, keeping capacity but never retaining caller maps.
• Tests: differential fuzz vs encoding/json (3M execs clean), cycle and depth-boundary regression tests, golden and escaping suites unchanged.

vet, test, and race pass; the 1M-line demos are byte-identical to main.